Responsible AI – Transparency and Fairness of data-based applications in practice

Organized by Eraneos, FHGR and ZHAW

Objectives

The usage of AI will soon be regulated in the European Union, with the upcoming AI Act which is currently in discussion and will be put into place in the very near future. Similar to the GDPR, this new EU regulation will impact the data science activities of Swiss companies dramatically: the use of many data-based algorithms, applications, and decision procedures will have to be re-thought, and often adopted, in order to address the requirements of the regulation.

In this workshop we will focus on two relevant requirements for AI systems which pose new challenges to data scientists:

  • Requirement of Explainability and Transparency of prediction, recommendation, and decision models
  • Requirement of Non-discrimination and Fairness of AI systems towards social groups

Distinguished speakers will explain these two requirements and comment on the state-of-the-art of how to implement such requirements technically, in a concrete data-science application.

In addition, participants will have the opportunity to discuss their specific challenges, open questions, etc.  with the experts as well as with the other workshop participants, in a moderated exchange format.

We expect that, with respect to practical implementation for real-world use cases, many questions may found to be still unexplored, requiring innovative approaches and further research to answer them. Representatives of the Databooster program will be present to support participants in receiving further support for such questions. This support will be delivered individually and specifically after the conference, but targeted innovation initiatives will be kickstarted on the spot already.

Agenda

Part 1: Non-discrimination and Fairness

In the first part of the workshop, we give an introduction into the field of algorithmic fairness, touching on questions such as:

  • What is meant by non-discrimination and fairness?
  • What kind of approaches exist to define and measure fairness of an AI system?
  • What are methods to make sure that data-based decision or recommendation algorithms still achieve maximum performance while satisfying a fairness constraint?

Part 2: Transparency

The second part of the work will be devoted to an introduction into transparency

  • What is meant by explainability and transparency?
  • What are methods to create transparency technically?
  • What are methods to make AI models explainable?

Break (30 min)

Part 3: Discussion and exchange, identification of open questions, kickstart innovation projects

In the third part of the workshop, we organize different exchange and discussion groups on challenges with respect to the practical implementation of solutions for addressing the two focus requirements. We will do this in an agile and moderated format, based on the needs of the participants.

Open questions and challenges for which no solutions seem yet to be available are identified. For those questions which qualify for an innovation support program, an innovation initiative may be kickstarted on the spot.

The Swiss Innovation Catalyst in data-driven value creation